A jewelry lookbook shot as one camera move

Ilvère is a jewelry brand I made up for this piece, and its Tidewater set is the whole site: a pair of studs, a pendant and a ring, each a green tourmaline in hand-shaped gold. The page opens on a full-length shot of the model wearing all three. Scroll, swipe or press an arrow key and the camera moves in one continuous take, up to her ear, down to the pendant, then to the ring on her hand. At each piece it stops, and that piece's product panel slides in.

I wanted to find out two things. Can a product page work like a short film the visitor steers? And can AI make both the film and the products without the two drifting apart?

The piece started as an adaptation of an interaction study: a looping video that zoomed from a model to her earring, where a product panel slid in, and then down to her ring. The loop played on its own, the ring got a close-up but no panel, and the ring in the product card wasn't the ring on her hand. I rebuilt the idea as a React component inside this site, with GSAP running the motion, and handed the camera to the visitor.

One number runs the page: how far along the take you are, from 0 at the full look to 3 at the ring. It sets the video's time, fades the opening copy, fills the index along the bottom and decides where a narrow screen crops the frame. Scrolling moves it directly. When you stop, it settles on the nearest piece in the direction you were scrolling, and it runs backwards too.

Every photo and every frame of video here was generated in Higgsfield. I designed the jewelry first, as three product shots made with Nano Banana 2.1, so the products would be the fixed reference for everything after them. Then came the model: one woman in an oversized ivory suit, wearing those exact pieces, in the hard light and dark setting of the original study. I made a still for each stop on the camera's path (the full look, her ear, her neckline and her hand) and checked every piece in every still against its product shot. Two went back. The first full-length shot gave the ring a diamond halo it doesn't have, and the first hand close-ups put notched lapels on a collarless jacket.

The video came last. Seedance 2.0 can start and end a clip on frames you give it, so each five-second clip runs from one approved still to the next. I tried the first clip on Seedance and on FLUX 3 Video and kept Seedance, because it landed on the end frame without a jump. The three clips join into one 15-second take, and the page loads it whole before it starts, so scrubbing never waits on the network.

These are the parts I designed on top of the original study:

• You drive the camera. Scroll, swipe or use the arrow keys and it stops at each piece, forwards or back.
• Every piece has a product panel, the ring included, and every product shot matches the piece on the model.
• Rings on the full look mark the pieces. Hover one for its name and price, or click it to send the camera straight there. The index along the bottom works the same way.
• On phones and narrow windows the frame slides to keep each piece in view, and on phones the panel becomes a bottom sheet.
• With reduced motion turned on, the camera cuts straight to each piece instead of moving, and a screen reader announces each piece and its price as it arrives.

The finished piece shops like a product page and plays like a short film, at whatever pace the visitor sets. It also came at the AI question from the other side to Atelier Sorène. There, each product was generated from the model's photo. Here the products came first and everything else was checked against them, which kept the film and the shop in agreement at every stop.